[llvm] [LV] Prefer fixed VFs for epilogue vectorization (PR #211814)
David Sherwood via llvm-commits
llvm-commits at lists.llvm.org
Fri Aug 7 01:58:22 PDT 2026
================
@@ -37,23 +34,26 @@ define void @simple_histogram(ptr noalias %buckets, ptr readonly %indices, i64 %
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_EXIT:%.*]], label [[VEC_EPILOG_ITER_CHECK:%.*]]
; CHECK: vec.epilog.iter.check:
-; CHECK-NEXT: [[MIN_EPILOG_ITERS_CHECK:%.*]] = icmp ult i64 [[N_MOD_VF]], [[TMP5]]
+; CHECK-NEXT: [[MIN_EPILOG_ITERS_CHECK:%.*]] = icmp ult i64 [[N_MOD_VF]], 2
; CHECK-NEXT: br i1 [[MIN_EPILOG_ITERS_CHECK]], label [[SCALAR_PH]], label [[VEC_EPILOG_PH]], !prof [[PROF3:![0-9]+]]
; CHECK: vec.epilog.ph:
; CHECK-NEXT: [[VEC_EPILOG_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[VECTOR_PH]] ]
-; CHECK-NEXT: [[TMP24:%.*]] = call i64 @llvm.vscale.i64()
-; CHECK-NEXT: [[TMP25:%.*]] = shl nuw i64 [[TMP24]], 1
-; CHECK-NEXT: [[N_MOD_VF2:%.*]] = urem i64 [[N]], [[TMP25]]
+; CHECK-NEXT: [[N_MOD_VF2:%.*]] = urem i64 [[N]], 2
; CHECK-NEXT: [[N_VEC3:%.*]] = sub i64 [[N]], [[N_MOD_VF2]]
; CHECK-NEXT: br label [[FOR_BODY:%.*]]
; CHECK: vec.epilog.vector.body:
; CHECK-NEXT: [[INDEX4:%.*]] = phi i64 [ [[VEC_EPILOG_RESUME_VAL]], [[VEC_EPILOG_PH]] ], [ [[INDEX_NEXT6:%.*]], [[FOR_BODY]] ]
; CHECK-NEXT: [[TMP18:%.*]] = getelementptr inbounds i32, ptr [[INDICES]], i64 [[INDEX4]]
-; CHECK-NEXT: [[WIDE_LOAD5:%.*]] = load <vscale x 2 x i32>, ptr [[TMP18]], align 4
-; CHECK-NEXT: [[TMP19:%.*]] = zext <vscale x 2 x i32> [[WIDE_LOAD5]] to <vscale x 2 x i64>
-; CHECK-NEXT: [[TMP20:%.*]] = getelementptr inbounds i32, ptr [[BUCKETS]], <vscale x 2 x i64> [[TMP19]]
-; CHECK-NEXT: call void @llvm.experimental.vector.histogram.add.nxv2p0.i32(<vscale x 2 x ptr> [[TMP20]], i32 1, <vscale x 2 x i1> splat (i1 true))
-; CHECK-NEXT: [[INDEX_NEXT6]] = add nuw i64 [[INDEX4]], [[TMP25]]
+; CHECK-NEXT: [[WIDE_LOAD5:%.*]] = load <2 x i32>, ptr [[TMP18]], align 4
+; CHECK-NEXT: [[TMP16:%.*]] = zext <2 x i32> [[WIDE_LOAD5]] to <2 x i64>
+; CHECK-NEXT: [[TMP17:%.*]] = extractelement <2 x i64> [[TMP16]], i64 0
----------------
david-arm wrote:
Something very odd is happening here. I don't see why we can't just also generate a
```
[[TMP20:%.*]] = getelementptr inbounds i32, ptr [[BUCKETS]], <2 x i64> [[TMP19]]
```
here. Any thoughts @huntergr-arm?
https://github.com/llvm/llvm-project/pull/211814
More information about the llvm-commits
mailing list